In the IT world, ‘Edge’ is understood as the endpoint or end device.
The reality is that this definition is very vague and can cover both complete systems that serve as a ‘gateway’ to other devices, as well as on-board elements of minimum consumption and very low cost.
What is clear is that in a world increasingly dominated by data, there is a tendency to provide advanced intelligence capabilities to these devices, the Edge Computing.
In this article we will look at three clear lines for the future in this regard.
Evolution of the concept of intelligence at the Edge
To understand these future lines, one must first look at how the concept of intelligence at the edge has evolved.
Due to low power, size and reliability requirements, early edge devices had extremely limited intelligence capabilities.
In general, they were isolated devices, applying very basic rules that were known in advance. Incredible as it may seem to us, most of the devices deployed in the world are confined to this first generation.
The emergence of the concept of cloud computing and, above all, the proliferation of low-cost and/or low-consumption communication technologies have led to the next generation: the connected device. This is the idea of extending the capabilities of the original devices by means of the intelligence available in the cloud (which has no limitations in terms of consumption or capacity), for which data is sent to this central node to make a more complex decision (the basic ones are still local) and act on the connected device.
But of course, this concept has obvious difficulties:
- The volume of data to be sent can be simply unmanageable (think video, but also thousands of time series), especially in very low power networks such as Sigfox, LoraWAN or Bluetooth LE.
- The time to make that decision is affected by the latency and delay imposed by these communication networks, and often the decision must be immediate.
So the third generation arises, that of the intelligent device that, in one way or another, will be able to make ‘advanced’ decisions on its own.
But of course, the power consumption, size, thermal profile and cost of the usual solutions for machine learning (GPUs) is directly unaffordable for most of these devices, unless they are of the ‘gateway’ type.
Intelligence trends at the Edge
Solutions of the Nvidia Jetson Nano or Google Coral type have been the queens so far in this field, making it possible to create intelligent gateway devices, with a profile close to embedded, with reasonable computational power and an affordable price if we think of a single device capable of acting as an intermediary for dozens of “dumb” devices. But these are getting a lot of pressure in three very specific scenarios, called to revolutionize the future of intelligence at the edge. Namely:
High computational power per W
Faced with the world of GPUs, interested only in raw power at any cost (to immortalize the presentation of Jen-Hsun Huang, CEO of Nvidia, taking his new GPU board out of a kitchen oven), new contenders are emerging interested in optimizing the maximum achievable computing power per watt of power consumption.
This allows us to access previously unknown levels of performance on the Edge while maintaining its thermal integrity and low power consumption. Essentially, like a Jetson Nano, but in a beastly way.
One of the most interesting in this regard, especially because of the agreements it is reaching with very important players, is Hailo. Solutions of this type bring the capabilities so far only available on the server to this type of device.
We are talking about the ability to run machine learning models on ultra-limited embedded devices.
We are talking about devices that cost a few euros, and of course have an extremely restricted hardware profile.
Despite these limitations, there are machine learning solutions focused on these types of devices, or even improved versions of these devices that incorporate accelerators for neural networks. Among the former, TinyML is a clear reference, with the company Edge Impulse as one of its best supporters. Among the latter, STMicroelectronics has just announced its line of ISPU microcontrollers.
Video is by far one of the largest consumers of bandwidth. It is therefore very difficult, if not impossible, to process it centrally on this type of device.
On the other hand, there are an infinite number of options for applying intelligence to a camera (think of the obvious case of security cameras, but also industrial process control).
In addition, the size of these images requires high computing power to decode and process them.
We will see a lot of evolution in this area, since, whether we like it or not, it benefits enormously from the advances in cell phones.
Acerca del autor
Consultant specialized in new technologies and Big Data.
Pioneer in Spain in the use of cutting-edge technologies such as Apache Kafka and Druid, he has extensive experience in the design of innovative technological products.